Connor Baker 7939de247b python3Packages.apex: 25.09 -> 26.09 ...

Diff: https://github.com/NVIDIA/apex/compare/25.09...26.09

- Drop fix-cuda-capabilities-selection.patch: upstream removed the hard-coded
  -gencode lists (NVIDIA/apex#1918) and the fmha/multihead_attn extensions
  (NVIDIA/apex#2016); every extension except group_norm_v2 (untouched by the
  patch before, too) now follows TORCH_CUDA_ARCH_LIST.
- Replace fix-cudnn-frontend-compat.patch, which needed regenerating for the
  reformatted sources, with a substituteInPlace of the one line it changed
  (21 identical occurrences in three files).
- Drop the lerp substitution, fixed upstream (NVIDIA/apex#1985).
- Only the cudnn-frontend submodule remains; neutralise its git calls.
- Drop cutlass and libcufile, used only by the removed multihead_attn and
  gpu_direct_storage extensions, and their import checks.
- tests/L0/run_transformer is gone with apex.transformer (NVIDIA/apex#1968).
- Bound build parallelism by NIX_BUILD_CORES: APEX_PARALLEL_BUILD extensions
  were built at once, each with ninja's default of one job per CPU of the
  machine (MAX_JOBS unset), times nvcc --threads 2. Build NIX_BUILD_CORES / 2
  extensions at once with MAX_JOBS=2, and drop --threads.

Assisted-by: Claude Code (Claude Opus 5.5)

2026-10-24 00:17:41 -07:00

5.3 KiB

Raw Permalink Blame History

View Raw